Improving Splenomegaly Segmentation by Learning from Heterogeneous Multi-Source Labels

Area of Science:

• Medical Imaging Analysis
• Deep Learning in Radiology
• Computational Anatomy

Background:

• Accurate splenomegaly segmentation in CT scans is crucial for diagnosing liver and spleen diseases.
• Manual annotation of medical imaging data is labor-intensive, leading to scarcity of labeled datasets.
• Existing deep learning models struggle with multi-source datasets containing heterogeneous labels.

Purpose of the Study:

• To develop a deep convolutional neural network (DCNN) method for splenomegaly segmentation using heterogeneous multi-resource labeled cohorts.
• To introduce a novel loss function that adaptively learns from multi-organ information across different datasets.
• To evaluate the proposed method's performance against established segmentation techniques.

Main Methods:

• A novel DCNN architecture integrating data from multiple sources with varying label sets.
• Introduction of a Dice similarity coefficient-based loss function for adaptive multi-organ learning.
• Experimental validation using three distinct CT scan cohorts with different labeling schemes.

Main Results:

• The proposed method achieved a median Dice similarity coefficient of 0.94 on the independent testing cohort.
• This performance was statistically superior (p-value<0.01) to multi-atlas segmentation (0.86), SS-Net (0.90), and U-Net (0.91).
• The adaptive loss function effectively leveraged heterogeneous labels for improved segmentation accuracy.

Conclusions:

• The developed co-learning strategy effectively integrates multi-source, heterogeneously labeled data for robust splenomegaly segmentation.
• The novel adaptive loss function demonstrates significant potential for training deep networks on diverse medical imaging datasets.
• This approach offers a promising solution for overcoming data scarcity in medical image analysis and is applicable beyond abdominal imaging.